SUMMER OF OPEN AI RESEARCH — ELEUTHERAI

0 Summer of Open AI Research Work on an AI research project under the mentorship of experienced researchers. Program Overview The Summer of Open AI Research is a 5-week, fully online research program sponsored by EleutherAI. We invite people with little research experience to contribute to open science under the mentorship of experienced researchers. We offer a selection of projects to apply to, listed below. Participants will be credited on their project, which may result in publication. Apply using the button to the right until June 8. We also encourage researchers who are willing to mentor a small group for five weeks to propose projects. Funding will be provided. Proposals are closed for the summer 2026 cohort. Who should apply? \- Experienced programmers interested in AI research \- BS/MS/PhD students in CS, math, physics, or related fields \- Self-taught researchers looking for structured mentorship \- Anyone wanting to contribute to open science! FAQ Q: Where is the event held? A: The EleutherAI Discord. Questions about the program should be asked there. Q: What topics/projects are offered? A: Check the projects list below. Q: Who is eligible to apply? A: Anyone! If you display that you have the ability to contribute, you will be considered. Q: Can I apply if I don’t have any research experience? A: Yes! In fact, we encourage you to do so. Part of SOAR’s goal is to give the opportunity to people outside of academia to get their first research experience. Q: How much time will I be expected to dedicate to SOAR? A: It will differ by project. Time estimates for each project are shown in the projects list below. Q: How competitive is it to become a mentee? What is the acceptance rate? A: We expect it to be quite competitive, depending on how many mentees each mentor takes. Q: Does “Open AI” have anything to do with the company behind ChatGPT? A: No, it means AI research done openly/collaboratively. See this paper for our research ethos. Apply here until June 8 Projects SearchTrackAllInterpretabilitySafetyApplications Interpretability Detecting Right-Answer, Wrong-Reason Behavior in Open-Weight Reasoning ModelsReasoningI-1Evaluating Interpretability Methods for Scientific ReasoningReasoningI-2Effective Rank and Semantic Compression in Foundation ModelsRepresentation LearningI-3How Do Foundation Models See the Same Sky?Representation LearningI-4Interpretable Conformal Prediction for Cross-Survey AstronomyRepresentation LearningI-5Does Structure Survive Scale? Diagnosing Hierarchy in SAEsMechanistic InterpretabilityI-6Replicating Seed Stability Results in Sparse AutoencodersMechanistic InterpretabilityI-7Activation Verbalizer Causality ExperimentsMechanistic InterpretabilityI-8Tracing Subliminal Learning to PretrainingMechanistic InterpretabilityI-9 Safety General Mechanisms Behind Subliminal PromptingAlignment & InterpretabilityS-1Predicting Inoculation Prompt Effectiveness from Latent Behavioral ShiftEmpirical AlignmentS-2Deceptive Compliance in LLM AgentsAgent EvaluationS-3 Applications AstroBridge: Connecting Multimodal Astronomical Observations to Open Language ModelsAI for ScienceA-1Scaling Information Retrieval for Domain Embedding ModelsInformation RetrievalA-2Editable Scene Simulation for Autonomous DronesComputer Vision & SimulationA-3Neolyre: Modern Singing Voice SynthesisAudio & Generative ModelingA-4 Interpretability / Reasoning Detecting Right-Answer, Wrong-Reason Behavior in Open-Weight Reasoning Models Mentor Pranava KumarMIT CSAIL Kellis Lab Participants 3-10 Time 6 hours/week This project studies whether open-weight reasoning models can arrive at correct answers for the wrong reasons, and whether interpretability tools can help detect that failure mode. Participants will create matched reasoning problems in clean, subtly hinted, and misleadingly hinted versions, then compare final answers, written reasoning, activations, and sparse-autoencoder features across conditions. The goal is to test whether internal evidence can distinguish genuine reasoning from shortcut-driven reasoning when surface behavior is misleading. Skills Participants should have Python experience, basic machine learning knowledge, and familiarity with PyTorch or Hugging Face. Participants should be comfortable with basic data analysis, Git/GitHub workflows, and careful reading of technical material. Careful experimental thinking is especially important. Prep Work • Read one short paper or post on chain-of-thought faithfulness. • Read an introduction to sparse autoencoders or mechanistic interpretability. • Set up Python with PyTor